CN104680235A - Design method of resonance frequency of circular microstrip antenna - Google Patents

Abstract

The invention discloses a design method of the resonance frequency of a circular microstrip antenna. The design method comprises the following steps: establishing mapping relationships between three relevant parameters, namely patch radius, dielectric substrate thickness and relative dielectric constant of the circular microstrip antenna, and the actually-measured resonance frequency by using a particle swarm neural network; by using a GPU technology, and in the unified computing device-framework programming environment, performing parallel accelerating computation on the training process of the particle swarm neural network, wherein the well-trained particle swarm neural network can be used for predicting the resonance frequencies of other circular microstrip antennae. By the design method, the defect that the computing time of the particle swarm neural network at a CPU end is long can be overcome, and the circular microstrip antenna resonance frequency modeling speed is increased and the circular microstrip antenna resonance frequency modeling accuracy can be improved.

Description

Circular Microstrip Antennas resonance frequency method for designing

Technical field

The present invention relates to a kind of Circular Microstrip Antennas resonance frequency method for designing, particularly relate to a kind of GPU based on unified calculation equipment framework and hold Parallel Particle Swarm Optimization neural network, to the method for designing of Circular Microstrip Antennas resonance frequency, belong to antenna technical field.

Background technology

Microstrip antenna is applied widely owing to having plurality of advantages, and wherein resonance frequency is an of paramount importance parameter in microstrip antenna designs process, directly determines the success or failure of Antenna Design.Neural network (Artificial Neural Networks, ANNs) model, owing to having good study and generalization ability, has been widely used in tie Microstrip Antenna design setting model.The neural network trained can set up mapping relations between microstrip antenna correlation parameter (comprising patch size, dielectric substrate thickness, relative dielectric constant) and actual measurement resonance frequency, thus the prediction completed other tie Microstrip Antenna, the method (being mainly divided into analytical method and numerical method two class) that the method is more traditional possesses clear superiority in precision, and effectively can overcome analytical method to the inapplicable shortcoming of many paster structures, and numerical method needs the shortcoming that recalculates to the change of paster geometry.Particle group optimizing (Particle Swarm Optimization) is as the global optimization approach of a kind of easy realization, fast convergence rate, just be applied to gradually in the training of neural network power threshold value, form so-called PSO Neural Network (Particle Swarm Optimization-Artificial Neural Networks, PSO-ANNs), can obtain better convergence precision and stronger predictive ability than conventional error backward propagation method, PSO Neural Network is also used to tie Microstrip Antenna modeling problem.

The PSO Neural Network of prior art exist when being used for tie Microstrip Antenna modeling a large problem be that the training time is longer, its reason is that complex network structures and a large amount of number of particles cause computation complexity higher.Utilize the concurrency of the natural individual in population behavior possessed in particle swarm optimization algorithm, adopting the parallelization of GPU technology to accelerate training PSO Neural Network is the effective thinking solving this problem.Therefore, a kind of GPU based on unified calculation equipment framework of research and design holds Parallel Particle Swarm Optimization neural network, and the rapid modeling tool being applied to Circular Microstrip Antennas resonance frequency is of great significance.

Summary of the invention

The object of the present invention is to provide a kind of Circular Microstrip Antennas resonance frequency method for designing, GPU based on unified calculation equipment framework holds Parallel Particle Swarm Optimization neural network, rapid modeling is carried out, to overcome the shortcoming of prior art CPU telomere subgroup neural computing overlong time to the design of Circular Microstrip Antennas resonance frequency.

Object of the present invention is achieved by the following technical programs:

A kind of Circular Microstrip Antennas resonance frequency method for designing, comprises the following steps:

Step 1: the neural network model building Circular Microstrip Antennas resonance frequency, by Circular Microstrip Antennas TM ₁₁the train samples of resonance frequency and test sample book normalized under pattern, each sample packages is containing these 4 data of paster radius, dielectric substrate thickness, relative dielectric constant and actual measurement resonance frequency; Determine the nodes of the input layer of neural network model, hidden layer and output layer, determine the hidden layer of neural network model and the activation function of output layer;

Step 2: the PSO Neural Network model building Circular Microstrip Antennas resonance frequency, is encoded into a D dimensional vector X by each particle _i, i represents particle numbering, i=1, and 2 ..., N, N are the number of particles in population, D dimensional vector X _irepresent the entitlement threshold value of a neural network, the output error quadratic sum of the normalization training sample of each neural network is the fitness value F (X of corresponding particle _i), the following parameters value in setting particle cluster algorithm: number of particles N, inertia weight w, Studying factors c ₁and c ₂, frequency of training T _max;

Step 3:CPU holds initialization PSO Neural Network: the position X of each particle of random initializtion _idwith speed V _id, d=1,2 ..., D, calculates the fitness value F (X of each particle _i), the individual optimal-adaptive angle value F (P of each particle _{i, best}) initial value is set to F (X _i), the personal best particle P of each particle _{id, best}initial value is set to X _id, all F (P _{i, best}) minimum value and the position of correspondence be set to global optimum fitness value F (G respectively _best) and global optimum position G _{d, best};

Step 4: carry out CPU and hold the transmission of GPU end data: CPU end calls cudaMemcpy () function, by the data X that CPU holds _id, V _id, F (X _i), P _{id, best}, F (P _{i, best}), G _{d, best}, F (G _best) reach GPU global memory; CPU end calls cudaMemcpyToSymbol () function, and the normalization number of training of being held by CPU is reportedly to GPU constant internal memory;

Step 5: carry out GPU and hold Parallel Particle Swarm Optimization neural computing: the concurrency utilizing the behavior of PSO algorithm individual in population, GPU holds the corresponding particle of a thread, repeatedly performs T at GPU end _maxsecondary Parallel Particle Swarm Optimization neural network algorithm iteration, each Parallel Particle Swarm Optimization neural network algorithm comprises following (4) four, (1) (2) (3) step that order successively performs:

(1) upgrade the speed V of each particle by the renewal of population speed and location updating formula simultaneously _id(t) and position X _id(t):

V _id(t+1)＝wV _id(t)+c ₁r ₁(P _id，best(t)-X _id(t))+c ₂r ₂(G _d，best(t)-X _id(t))

X _id(t+1)＝X _id(t)+V _id(t+1)

In formula, current iteration number of times t=1,2 ..., T _max, r ₁and r ₂be the equally distributed random number between [0,1], the curand_uniform () function in CURAND storehouse can be used in GPU end generation random number);

(2) calculate the fitness value F (X that each particle is corresponding simultaneously _i);

(3) upgrade the individual optimal-adaptive angle value F (P of each particle simultaneously _{i, best}) and the personal best particle P of correspondence _{id, best}if: F (X _i) < F (P _{i, best}), then F (X _i)=F (P _{i, best}), P _{id, best}=X _id;

(4) global optimum fitness value F (G is upgraded with parallel reduction algorithm _best) and the global optimum position G of correspondence _{d, best}if: min (F (P _{i, best})) < F (G _best), then F (G _best)=min (F (P _{i, best})), i=I, G when getting min _{d, best}=P _{id, best};

Step 6: carry out GPU and hold the transmission of CPU end data, namely CPU end calls cudaMemcpy () function, GPU is held the neural network optimum power threshold value G trained _{d, best}be transmitted back to CPU end;

Step 7: normalized training sample and test sample book are brought into the neural network trained, exports anti-normalizing by network, obtains the network output valve of Circular Microstrip Antennas resonance frequency.

Object of the present invention can also be realized further by following technical measures:

Aforementioned circle tie Microstrip Antenna method for designing, wherein the nodes of the input layer of neural network described in step 1, hidden layer and output layer is:

Neural network adopts 3 conventional layer network structures, and network input layer nodes i is 3, output layer nodes j is 1, and the span of the number of hidden nodes p is determined by following formula:

$p=\sqrt{i+j}+\mathrm{\&alpha;},1\&le;\mathrm{\&alpha;}\&le;10;$

Wherein the hidden layer of neural network described in step 1 and the activation function of output layer are determined as follows:

The hidden layer activation function of neural network elects bipolarity S type function as, is shown below:

$f\left(u\right)=\frac{2}{1+\exp (-2\&times;u)}-1,-\&infin;<u<+\&infin;;$

The activation function of output layer elects unipolarity S type function as, is shown below:

$f\left(u\right)=\frac{1}{1+\exp (-u)},-\&infin;<u<+\&infin;.$

Aforementioned circle tie Microstrip Antenna method for designing, wherein described in step 1, the number of hidden nodes p is set as 8.

Aforementioned circle tie Microstrip Antenna method for designing, wherein vectorial X described in step 2 _iparticle dimension D is 41.

Aforementioned circle tie Microstrip Antenna method for designing, wherein number of particles N described in step 2 is the multiple value of 64; Inertia weight w value is 1 to 0.4 linear decrease; Studying factors c ₁and c ₂get 2.8 and 1.3 respectively; Frequency of training T _maxvalue is 1000.

Compared with prior art, the invention has the beneficial effects as follows:

(1) by holding the computation process of each particle of parallel processing to achieve at GPU, the parallel accelerate of PSO Neural Network is trained, the modeling time of remarkable minimizing Circular Microstrip Antennas resonance frequency, the highest calculating speed-up ratio obtaining 347 times under the prerequisite that optimizing stability is consistent; (2) population is more, and the speed-up ratio of acquisition is higher; (3) the present invention significantly increases population is the specific process adapting to GPU computing architecture; Along with the increase of population, compared with holding serial PSO Neural Network with prior art CPU, it is very limited that GPU holds increase the working time of Parallel Particle Swarm Optimization neural network; If significantly increase population at GPU end, the method significantly can reduce modeling error under the modeling time increases extremely limited situation, and modeling error performance is better than the effect of all prior aries.

Accompanying drawing explanation

Fig. 1 is Circular Microstrip Antennas illustraton of model;

Fig. 2 is unified calculation equipment framework programming model figure;

Fig. 3 is Circular Microstrip Antennas resonance frequency method for designing process flow diagram of the present invention;

Fig. 4 is Circular Microstrip Antennas TM ₁₁the train samples of resonance frequency and test sample book under pattern;

Fig. 5 is the population vector coding model of 3-8-1 artificial neural;

Fig. 6 is the computing equipment platform configuration figure that the present invention uses;

Fig. 7 is that GPU holds Parallel Particle Swarm Optimization neural network to Circular Microstrip Antennas TM ₁₁the calculating speed-up ratio data that under pattern, modelling resonance frequencies obtains;

Fig. 8 is the Circular Microstrip Antennas TM that in existing document, various CPU terminal nerve network model obtains ₁₁mean absolute error sum data under pattern between the network output valve of resonance frequency and measured value.

Embodiment

The present invention is directed to PSO Neural Network to long problem computing time during Circular Microstrip Antennas modelling resonance frequencies, hold on the basis of Parallel Particle Swarm Optimization research at existing GPU, design and Implement a kind of GPU based on unified calculation equipment framework and hold Parallel Particle Swarm Optimization Artificial Neural Network, and rapid modeling has been carried out to Circular Microstrip Antennas resonance frequency, modeling speed and modeling accuracy are got a promotion.

Below in conjunction with the drawings and specific embodiments, the invention will be further described.

Be illustrated in figure 1 a Circular Microstrip Antennas model, the radius of circular paster 1 is a, and the thickness of dielectric substrate 2 is h, and relative dielectric constant is ε _r.Cavity model method is adopted to analyze this circular patch, as h < < λ ₀time, the cavity between paster 1 and ground plate 3 can be regarded surrounding as and be magnetic wall, is the thin cylinder resonance chamber of electric wall up and down, and the electric field in chamber should for only having E _ztransverse magnetic wave (Transverse-magnetic Wave, the TM ripple) pattern of component.By E _zexpand into the superposition of eigenfunction:

$E_z=\underset{m=0}{\overset{\&infin;}{\mathrm{\&Sigma;}}}\underset{n=0}{\overset{\&infin;}{\mathrm{\&Sigma;}}}{C}_{\mathrm{mn}}{\mathrm{\&Psi;}}_{\mathrm{mn}}---\left(3\right)$

Eigenfunction and boundary condition should have following form:

J _m' (k _mna) J in=0 (5) formula _m(k _mnρ) be m rank Bessel's function, k _mnfor separation parameter, and

$k_{\mathrm{mn}}=\frac{{{\mathrm{\&chi;}}^{\&prime;}}_{\mathrm{mn}}}{a}---\left(6\right)$

χ ' in formula _mnfor function J _mn-th of ' (χ).By k=k _mnthe resonance frequency obtaining Circular Microstrip Antennas can be similar to

$f_{\mathrm{mn}}=\frac{{{\mathrm{c\&chi;}}^{\&prime;}}_{\mathrm{mn}}}{2\mathrm{\&pi;a}\sqrt{{\mathrm{\&epsiv;}}_r}}---\left(7\right)$

In formula, c is the velocity of propagation of electromagnetic wave at free space, m and n is integer.When calculating the holotype TM of circular patch ₁₁during the resonance frequency of mould, formula (7) can be write as

$f_{11}=\frac{0.293c}{a\sqrt{{\mathrm{\&epsiv;}}_r}}---\left(8\right)$

If consideration edge effect, the equivalent redius a that the radius a in formula can calculate by following experimental formula _ereplace:

$a_e=a{[1+\frac{2h}{{\mathrm{\&pi;a\&epsiv;}}_r}(\ln \frac{\mathrm{\&pi;a}}{2h}+1.7726)]}^{\frac{1}{2}}---\left(9\right)$

Obvious, Circular Microstrip Antennas resonance frequency depends on a, h, ε _r, m, n; At TM ₁₁a is depended on, h, ε under pattern _r.

Be illustrated in figure 2 unified calculation equipment framework programming model, it is using CPU as main frame, and GPU is as coprocessor, and both collaborative works, Each performs its own functions.CPU is responsible for carrying out the strong issued transaction of logicality and serial computing, and GPU is then absorbed in the highly threading parallel processing task of execution.Unified calculation equipment framework adopts single instrction multithreading (Single Instruction Multiple Threads, SIMT) execution pattern.Kernel (kernel) function performs the parallel computation task on GPU, is a step that can be executed in parallel in whole program.Sets of threads is made into the different level of block grid (Grid), thread block (Block), thread (Thread) these three by unified calculation equipment framework, and adopt multi-level memory construction: only to the visible local storage of single thread, to the visible shared storage of thread in block, to the visible global storage of all threads etc.The parallel of two levels is there is, the coarse grain parallelism not needing to communicate between the thread block namely in same grid, the fine grained parallel of the permission communication between the thread in same thread block in a kernel function.Unified calculation equipment framework calculation process generally includes the transmission of CPU to GPU data, kernel function performs, GPU transmits three steps to cpu data.

The GPU be illustrated in figure 3 based on unified calculation equipment framework holds Parallel Particle Swarm Optimization neural network to Circular Microstrip Antennas TM ₁₁under pattern, the algorithm flow chart of rapid modeling is carried out in resonance frequency design, and concrete steps are as follows:

Step 1: the neural network model building Circular Microstrip Antennas resonance frequency, comprises Circular Microstrip Antennas TM ₁₁the train samples of resonance frequency and test sample book (each sample packages is containing these 4 data of paster radius, dielectric substrate thickness, relative dielectric constant and actual measurement resonance frequency) normalized under pattern, determine the nodes of the input layer of neural network, hidden layer and output layer, determine the hidden layer of neural network and the activation function of output layer.

(1) by Circular Microstrip Antennas TM ₁₁the train samples of resonance frequency and test sample book (each sample packages is containing these 4 data of paster radius, dielectric substrate thickness, relative dielectric constant and actual measurement resonance frequency) normalized under pattern.

16 groups of training samples (not being with asterisk) and 4 groups of test sample books (band asterisk) are listed, wherein the paster radius of Circular Microstrip Antennas, dielectric substrate thickness, relative dielectric constant these 3 correlation parameters (a, h, ε in Fig. 4 _r) as network input (the 2 to the 4 row of Fig. 4), at TM ₁₁actual measurement resonance frequency f corresponding under pattern _mEexport (the 5th row of Fig. 4) as network.For convenience of data processing, every column data is all done normalized, and restricting data scope is in [0,1].Normalized training sample is used during neural metwork training.

(2) nodes of the input layer of neural network, hidden layer and output layer is determined.

Neural network adopts 3 conventional layer network structures, and network input layer nodes i is 3 (corresponding a, h, ε respectively _r), output layer nodes j is 1 (respective resonant frequencies), the span of the number of hidden nodes p is determined by following formula:

$p=\sqrt{i+j}+\mathrm{\&alpha;},1\&le;\mathrm{\&alpha;}\&le;10---\left(10\right)$

Preferably, the number of hidden nodes p is set as 8, such neural network structure is 3-8-1.

(3) hidden layer of neural network and the activation function of output layer is determined.

Elect the hidden layer activation function of neural network as bipolarity S type function according to problematic features, its expression formula is such as formula shown in (11); The activation function of output layer elects unipolarity S type function as, and its expression formula is such as formula shown in (12).

$f\left(u\right)=\frac{2}{1+\exp (-2\&times;u)}-1,-\&infin;<u<+\&infin;---\left(11\right)$

$f\left(u\right)=\frac{1}{1+\exp (-u)},-\&infin;<u<+\&infin;---\left(12\right)$

Step 2: the PSO Neural Network model building Circular Microstrip Antennas resonance frequency, comprises and each particle is encoded into a D dimensional vector X _i(i represents particle numbering, i=1,2, ..., N, N is the number of particles in population) represent the entitlement threshold value of a neural network, the output error quadratic sum (Sum of Squared Error, SSE) of the normalization training sample of each neural network is the fitness value F (X of corresponding particle _i); Following parameters value in setting particle cluster algorithm: number of particles N, inertia weight w, Studying factors c ₁and c ₂, frequency of training T _max.

(1) each particle is encoded into a D dimensional vector X _i(i represents particle numbering, i=1,2, ..., N, N is the number of particles in population) represent the entitlement threshold value of a neural network, the output error quadratic sum (Sum of Squared Error, SSE) of the normalization training sample of each neural network is the fitness value F (X of corresponding particle _i).

Fig. 5 is the population vector coding model of 3-8-1 artificial neural, this model adopts particle respectively to tie up and respectively weighs threshold value principle one to one with neural network, each particle is encoded into the vectorial X that a particle dimension (problem dimension, power threshold number) D is 41 _i(i represents particle numbering, i=1, and 2 ..., N, N are the number of particles in population) represent the entitlement threshold value of a neural network, shown in (13); The output error quadratic sum (Sum of Squared Error, SSE) of the normalization training sample of each neural network is the fitness value F (X of corresponding particle _i).

X _i＝[X _i1X _i2X _i3…X _i39X _i40X _i41] (13)

(2) the following parameters value in particle cluster algorithm is set: number of particles N, inertia weight w, Studying factors c ₁and c ₂, frequency of training T _max.

In unified calculation equipment framework, a thread bundle (Warp) comprises 32 adjacent threads of index, stream multiprocessor (Stream Multiprocessor, SM) scheduling and execution thread in units of thread bundle, therefore number of particles N is designed to the multiple value of 32, again because number of particles generalized case is greater than problem dimension 41, therefore number of particles N is designed to the multiple value of 64; Inertia weight w value 1 to 0.4 linear decrease; Studying factors c ₁and c ₂get 2.8 and 1.3 respectively; Frequency of training T _maxvalue 1000.

Step 3:CPU holds initialization PSO Neural Network, comprises the position X of each particle of random initializtion _idwith speed V _id(d=1,2 ..., D), calculate the fitness value F (X of each particle _i), the individual optimal-adaptive angle value F (P of each particle _{i, best}) initial value is set to F (X _i), the personal best particle P of each particle _{id, best}initial value is set to X _id, all F (P _{i, best}) minimum value and the position of correspondence be set to global optimum fitness value F (G respectively _best) and global optimum position G _{d, best}.

Step 4:CPU holds the transmission of GPU end data, comprises CPU end and calls cudaMemcpy () function, by the particle data X that CPU holds _id, V _id, F (X _i), P _{id, best}, F (P _{i, best}), G _{d, best}, F (G _best) reach GPU global memory; CPU end calls cudaMemcpyToSymbol () function, and the normalization number of training of being held by CPU is reportedly to GPU constant internal memory.

Step 5:GPU holds Parallel Particle Swarm Optimization neural computing, comprises the concurrency utilizing the behavior of PSO algorithm individual in population, and GPU holds the corresponding particle of a thread, repeatedly performs T at GPU end _maxsecondary Parallel Particle Swarm Optimization neural network algorithm iteration, each Parallel Particle Swarm Optimization neural network algorithm comprises following (1) (2) (3) (4) four steps that order successively performs:

(1) upgrade the speed V of each particle by the renewal of population speed and location updating formula simultaneously _id(t) and position X _id(t);

V _id(t+1)＝wV _id(t)+c ₁r ₁(P _id，best(t)-X _id(t))+c ₂r ₂(G _d，best(t)-X _id(t)) (1)

X _id(t+1)＝X _id(t)+V _id(t+1) (2)

In formula, current iteration number of times t=1,2 ..., T _max, r ₁and r ₂it is the equally distributed random number (the curand_uniform () function in CURAND storehouse can be used in GPU end generation random number) between [0,1];

(2) calculate the fitness value F (X that each particle is corresponding simultaneously _i);

(3) upgrade the individual optimal-adaptive angle value F (P of each particle simultaneously _{i, best}) and the personal best particle P of correspondence _{id, best}if: F (X _i) < F (P _{i, best}), then F (X _i)=F (P _{i, best}), P _{id, best}=X _id;

(4) global optimum fitness value F (G is upgraded with parallel reduction algorithm _best) and the global optimum position G of correspondence _{d, best}if: min (F (P _{i, best})) < F (G _best), then F (G _best)=min (F (P _{i, best})) when min (get i=I), G _{d, best}=P _{id, best}.

Step 6:GPU holds the transmission of CPU end data, comprises CPU end and calls cudaMemcpy () function, GPU is held the neural network optimum power threshold value G trained _{d, best}be transmitted back to CPU end.

Step 7: normalized training sample and test sample book are brought into the neural network trained, exports anti-normalizing by network, obtains the network output valve of Circular Microstrip Antennas resonance frequency.Obtain the absolute error summation of network output valve and measured value, the T.T. that logging program runs, can performance evaluation be done to modeling speed and modeling error.

Under facility environment shown in Fig. 6, carry out numerical value test according to above-mentioned steps, result as shown in Figure 7." calculating speed-up ratio " item in Fig. 7, as conventional acceleration index, is defined as PSO Neural Network CPU under identical population with identical iterations (getting 1000 times in experiment) and holds program runtime and GPU to hold the ratio of program runtime.Fig. 8 lists the Circular Microstrip Antennas TM that in existing document, various CPU terminal nerve network model obtains ₁₁mean absolute error summation under pattern between the network output valve of resonance frequency and measured value, contrasts with result of calculation of the present invention to facilitate.

Following analysis is done to Fig. 7 and Fig. 8:

A () population is more, the calculating speed-up ratio of acquisition is higher, and GPU holds the highest speed-up ratio obtaining 347 times of Parallel Particle Swarm Optimization neural network.Double along with population, when population is less than or equal to 16384 (the maximum resident Thread Count of GPU used is 26624), calculates speed-up ratio roughly double; When population is more than or equal to 32768, calculating speed-up ratio still can increase but slowdown in growth.

B () GPU holds Parallel Particle Swarm Optimization neural network to have the optimizing stability of holding serial PSO Neural Network same with CPU.Along with being on the increase of population, the error of CPU program and GPU program constantly reduces; When population is identical, CPU program is roughly the same with the error of GPU program or close.

C () significantly increases population is the specific process adapting to GPU computing architecture.Along with the increase of population, compared with holding with CPU, the working time of GPU end increases very limited.GPU holds the modeling error of Parallel Particle Swarm Optimization neural network, when population is more than or equal to 256, is better than the result of BP in document; When population is more than or equal to 1024, be better than the result of DBD in document; When population is more than or equal to 8192, be better than the result of PTS in document; When population is more than or equal to 32768, be better than the result of EDBD in document; When population is more than or equal to 65536, be better than comprising the result of all existing document of BiPSO in document.

Following summary is done to above analysis:

(1) if GPU end uses and CPU holds identical population, the method can significantly reduce the modeling time under the conforming prerequisite of convergence.

(2) if significantly increase population at GPU end, the method significantly can reduce modeling error under the modeling time increases extremely limited situation.

The method significantly can reduce modeling error under the modeling time increases extremely limited situation, and modeling error performance is better than the effect of all prior aries.

In addition to the implementation, the present invention can also have other embodiments, and all employings are equal to the technical scheme of replacement or equivalent transformation formation, all drop in the protection domain of application claims.

Claims (5)

1. a Circular Microstrip Antennas resonance frequency method for designing, is characterized in that, the method comprises the following step;

Step 1: the neural network model building Circular Microstrip Antennas resonance frequency, by Circular Microstrip Antennas TM ₁₁the train samples of resonance frequency and test sample book normalized under pattern, each sample packages is containing these 4 data of paster radius, dielectric substrate thickness, relative dielectric constant and actual measurement resonance frequency; Determine the nodes of the input layer of neural network model, hidden layer and output layer, determine the hidden layer of neural network model and the activation function of output layer;

Step 2: the PSO Neural Network model building Circular Microstrip Antennas resonance frequency, is encoded into a D dimensional vector X by each particle _i, i represents particle numbering, i=1, and 2 ..., N, N are the number of particles in population, D dimensional vector X _irepresent the entitlement threshold value of a neural network, the output error quadratic sum of the normalization training sample of each neural network is the fitness value F (X of corresponding particle _i), the following parameters value in setting particle cluster algorithm: number of particles N, inertia weight w, Studying factors c ₁and c ₂, frequency of training T _max;

Step 3:CPU holds initialization PSO Neural Network; The position X of each particle of random initializtion _idwith speed V _id, d=1,2 ..., D, calculates the fitness value F (X of each particle _i), the individual optimal-adaptive angle value F (P of each particle _{i, best}) initial value is set to F (X _i), the personal best particle P of each particle _{id, best}initial value is set to X _id, all F (P _{i, best}) minimum value and the position of correspondence be set to global optimum fitness value F (G respectively _best) and global optimum position G _{d, best};

Step 4: carry out CPU and hold the transmission of GPU end data: CPU end calls cudaMemcpy () function, by the data X that CPU holds _id, V _id, F (X _i), P _{id, best}, F (P _{i, best}), G _{d, best}, F (G _best) reach GPU global memory; CPU end calls cudaMemcpyToSymbol () function, and the normalization number of training of being held by CPU is reportedly to GPU constant internal memory;

Step 5: carry out GPU and hold Parallel Particle Swarm Optimization neural computing: the concurrency utilizing the behavior of PSO algorithm individual in population, GPU holds the corresponding particle of a thread, repeatedly performs T at GPU end _maxsecondary Parallel Particle Swarm Optimization neural network algorithm iteration, each Parallel Particle Swarm Optimization neural network algorithm comprises following (4) four, (1) (2) (3) step that order successively performs:

(1) upgrade the speed V of each particle by the renewal of population speed and location updating formula simultaneously _id(t) and position X _id(t):

V _id(t+1)＝wV _id(t)+c ₁r ₁(P _id,best(t)-X _id(t))+c ₂r ₂(G _d,best(t)–X _id(t))

X _id(t+1)＝X _id(t)+V _id(t+1)

In formula, current iteration number of times t=1,2 ..., T _max, r ₁and r ₂be the equally distributed random number between [0,1], the curand_uniform () function in CURAND storehouse can be used in GPU end generation random number);

(2) calculate the fitness value F (X that each particle is corresponding simultaneously _i);

(3) upgrade the individual optimal-adaptive angle value F (P of each particle simultaneously _{i, best}) and the personal best particle P of correspondence _{id, best}; If F is (X _i) <F (P _{i, best}), then F (X _i)=F (P _{i, best}), P _{id, best}=X _id;

(4) global optimum fitness value F (G is upgraded with parallel reduction algorithm _best) and the global optimum position G of correspondence _{d, best}if: min (F (P _{i, best})) <F (G _best), then F (G _best)=min (F (P _{i, best})), i=I, G when getting min _{d, best}=P _{id, best};

Step 6: carry out GPU and hold the transmission of CPU end data, namely CPU end calls cudaMemcpy () function, GPU is held the neural network optimum power threshold value G trained _{d, best}be transmitted back to CPU end;

Step 7: normalized training sample and test sample book are brought into the neural network trained, exports anti-normalizing by network, obtains the network output valve of Circular Microstrip Antennas resonance frequency.

2. Circular Microstrip Antennas resonance frequency method for designing as claimed in claim 1, is characterized in that, the nodes of the input layer of neural network described in step 1, hidden layer and output layer is:

Neural network adopts 3 conventional layer network structures, and network input layer nodes i is 3, output layer nodes j is 1, and the span of the number of hidden nodes p is determined by following formula:

$\begin{array}{cc}p=\sqrt{i+j}+\mathrm{\&alpha;}& 1\&le;\mathrm{\&alpha;}\&le;10\end{array};$

The hidden layer of neural network described in step 1 and the activation function of output layer are determined as follows:

The hidden layer activation function of neural network elects bipolarity S type function as, is shown below;

$\begin{array}{cc}f\left(u\right)=\frac{2}{1+\exp (-2\&times;u)}-1& -\&infin;<u<+\&infin;\end{array};$

The activation function of output layer elects unipolarity S type function as, is shown below;

$\begin{array}{cc}f\left(u\right)=\frac{1}{1+\exp (-u)}& -\&infin;<u<+\&infin;\end{array}.$

3. Circular Microstrip Antennas resonance frequency method for designing as claimed in claim 2, it is characterized in that, described the number of hidden nodes p is set as 8.

4. Circular Microstrip Antennas resonance frequency method for designing as claimed in claim 3, is characterized in that, vectorial X described in step 2 _iparticle dimension D is 41.

5. Circular Microstrip Antennas resonance frequency method for designing as claimed in claim 4, it is characterized in that, number of particles N described in step 2 is the multiple value of 64; Inertia weight w value is 1 to 0.4 linear decrease; Studying factors c ₁and c ₂get 2.8 and 1.3 respectively; Frequency of training T _maxvalue is 1000.